1 Introduction

Dualistic Algebra for Qualitative Analysis

Frangois Guerrin

Institut National de la Recherche Agronomique (INRA) - Biometrics and Artificial Intelligence UnitBP 27 - 31326 Castanet-Tolosan Cedex (F)phone +33 6128 52 88

fax +33 6128 53 35email guerrin@toulouse.inra.fr

Abstract: A central point in qualitative modelsis that of calculation . The problem is to repre-sent, compare, and combine, approximate orsymbolic values with the major constraint ofconsistency with some numerical counterpart .Aware of the fact that many qualitative characte-rizations are in the form of antagonistic concepts(such as low vs. high, weak vs. strong, bad vs .good, etc.) more or less explicitly separated by areference (we call the norm), we exploit this ideato build a dualistic algebra based on a linear or-der and an homomorphism with the set of inte-gers Z, which accounts for the ranks of qualita-tive values (viewed as labels expressing magni-tudes) ordered with respect to the norm. We de-fine operations on those values, globally consis-tent with classical arithmetic (+, -, x, /, power,root) as well as fitting common sense in reaso-ning .

The major goal of qualitative reasoning (QR) isto provide engineers with computational theoriesto allow them to draw inferences about systems,using information of various origins and diffe-rent natures . Due to the complexity of systems,qualitative models should be able to involve va-riables having possibly different domains suchas real numbers, numerical intervals or even pu-rely symbolic elements . Measurements (nume-rical values) as well as basic observations(symbolic values) are often translated into moremeaningful concepts within the framework of thesystem under study : 0.476 is far above normal,150 is very low, or small is beautiful. Conceptsused to characterize quantities are often antago-nistic : low/high, good/bad, above/below,false/true . This dualism implies the existence of areference representing normality within thecontext of reasoning. It is either explicitly statedor implicit. Actually it corresponds to the idea ofbeing located between both antagonistic conceptswhich yields an order w.r.t . the norm. It may benecessary to go beyond a binary opposition andcharacterize a quantity at different granularity,i.e . distinguishing among good and bad concepts

more subtle characterizations such as very goodor very very bad. Another big point is composi-tionality, i.e . how to combine very high withbad, below with average and even blue withbeautiful and cold! These needs of comparisonand combination of heterogeneous things lead usto find some set to play the role of a "commoncurrency". Because qualitative values are intrin-sically discreet entities, we chose the integers Z.Since we do not base principally our formalismon a partition of the real line (as opposed to mostof QR approaches), we will refer to qualityspaces (in the sense of Hayes, 1985) to denotethe set of possible values of a given qualitativevariable .This paper addresses three main aspects : (i) howto build a quality space (QS) with a dualisticstructure from a set of labels (section 2) ; (ii) howto compare values of (single or distinct) variablesexpressed on different QS at varying granularity(section 3) ; (iii) how to combine values throughoperations, consistent with both conventionalarithmetic and common sense expectations(section 4) . Then, we provide two examples ofqualitative analysis (section 5) . Finally we dis-cuss (too briefly) this so called dualistic algebra(DuAl) formalism in the light of quantity spacestheory (section 6) .

2 Quality Spaces

2 .1 Dualistic Variable

Def. 2.1 : A dualistic variable is a surjectivefunction x: D� -~ QS (x), with domain the codo-main D� of any variable (e.g . measurement, ob-servation, or even another dualistic variable) andcodomain a quality space QS(x), that is a finiteset of ordered labels, i .e . qualitative values, asdefined below .Let L(x) be a finite non-empty set of labels ofany kind

. .., Xw such that each of them isassumed to be a possible value for a variable x,and all of them cover the whole codomain of x.Let us assume they can be totally ordered (byany kind of objective or even arbitrary means)

according to a precedence relation

Thus,(L(x), < ) as a linear order, is reflexive, anti-symmetric, transitive, and obeys the dichotomylaw

XV E L(x), either Xt - 1), leads to consider 2n + 1possible qualitative values, i.e . an odd number.In our QS, if n is odd, the norm is a landmark ; ifn is even, the norm is an interval between twolandmarks.

P.2.5.1 : (QS (x), Z), linear order, has leastand greatest elements, gmin and gmax respecti-vely : `d qi E QSn(x), gmin L qi L gznax .For sake of clarity, we may write extensively aQS as: Igmin, . . ., qi, . . ., gmax1, where the middleelement is the norm qnm (e.g . in { small, medium,tall), medium is the norm) .

Example 2.5.1 : QS for a measurable variable .Let (0, 100) be the landmarks for the tempera-ture ofwater . 0°C, freezing point, and 100°C,boiling point. An appropriate description space,covering the whole variable's range, may bechosen as: QS*(water temp) = {(-273, 0), 0,(0, 100), 100, (100, +-)), where -273°C is theabsolute zero. Its elements could also be denotedas labels corresponding to water states, e.g . :QS'*(water temp) = {ice, freezing, liquid, boi-ling, vapor) . If one takes the liquid state("liquid" or (0, 100)) as norms, then both QS*and QS'* are quality spaces for water temp.They may be merged also within a single qualityspace, using equivalence between temperaturesand states : QS(water temp) = { {(-273, 0), ice),(0, freezing), [(0, 100), liquid), (100, boi-ling), ((100, +-), vapor)), where ((0, 100),liquid) is taken as the norm.

Example 2.5.2: QS for a binary variable." L(bits) = ( 1, no, true, zero, yes, one, 0, false)" QS*(bits) = HO, no, zero, false), (1, yes,one, true) ) is not a QS unless we add a norm toit, arbitrarily or not, depending on the granularityat which the phenomenon is observed (see howWilliams, 1984, represents a digital signal as a 5valued set) ; e.g . let us take the symbol ?.Assuming 0 L ? L 1, we write : QS(bits) = { (0,no, zero, false), ( ?), ( 1, yes, one, true)) .

Example 2.5.3 : QS for a non-measurablevariable.It may seem obvious that values of any measu-rable variable can be ordered. Let us now consi-der color with possible values: yellow, green andred; how linearly order these labels ? For this,we must defme more precisely what do we meanby color . We give hereafter several QS, depen-ding on the definition given to color as:" my favorite color for cars : (yellow, green, red)" increasing wavelength ranges in the light spec-

trum : (green, yellow, red)" traffic-light spots enumerated top-down : {red,yellow, green)" denoting the freshness of maple-tree leaves:(yellow, red, green)" my increasing preference to paint my dining-room : Ired, green, yellow)" the lexical order of their names: (green, red,yellow) .In every case the middle element has the meaningof being intermediate between the lowest and thehighest preference. Of course, one must considereach characterization as alternative choices : colorbased on different QS are distinct variables,since the QS is what makes the variable's se-mantics, not the converse .

2 .6 Distance

Def. 2.6: The distance between two QS ele-ments is a metric function with integers codo-main d: QS(x) x QS(x) -+ Z, defined for anypair (qi L 9j) by:" d(qi, (: j) = 0, iff qi = qi" d(qi, (:V) = +1, iff: -,3 qk s.t . qi L qk L qj" d(qi, qi) = d(qi, qi , ) + d(qi ,, qi),

such that d(qi,, Q = 1" d(qi, qj) = -d(qj, qi)

P. 2.6.1 : Composition of distancesIf: qi L qk (qi, qk E QS(x)), then :

dqj E QS(x) d(qi, qk) = d(qi, qi) + d(qi, qk)

P. 2.6.2: OrderVqi, qi E QS(x), qi L qi, iff: d(qi, qi) ? 0.

P. 2.6.3: QS cardinalityIQS(x)I = d(gmin, gmax) +1

P. 2.6.4: The norm is the middle element of aQS : d(gmin, qm) = d(qm, gmax) .

2 .7 Ranking Function of QS Elements

Def. 2.7: The ranking function of a qualityspace, R; QS(x) -4 Z is defined by:d qi E QS(x),Ngi) = d(qm, qi)

for Z the integers and qm the norm of QS (x) .Note that w.r.t. the definition :

Nqrn) = d(qm, qm) = 0.In the following we will denote a quality spaceas QSn(x) with n = R(gmax) = -R(grnin), and willcall the image of R. (also denoted &) the rankspace associated to QSn(x) .

P.2.7.1 :

R,is order-preserving, i.e .`d qi, qj E Qsn(x), qi L qi =:~ Nqi) :5 Nqi).

P.2.7.2: R, is injective .

2.8 Interpretation Function

Def. 2.8: Let Z denote the integers and QSn(x)be a quality space ; we define an interpretation asa dualistic variable I-. Z -> QSn(x), by: b` z E Z

q�. if z>nI(z) = J qi such that : R,(qi) = z, if - n :5 z :5 n

q,,,In if z < -n(I may also be denoted in) . By Def. 2.7, since

R(qm) = 0, then : I(0) = q.-

P.2 .8.1 : I is order-preserving, i.e .b' zi, zj E Z, zi

P.3.1.1 : Hierarchical inclusion .Given QSn(x) and QS,,~y), and their rank spacesRS,(x) and RSn~y), the lower-dimensional rankspace is included within the larger one:

n :5 n'=* RSn(x) C_ RSn~A

P.3.1.2 : Given QSn(x) and QSn'(x), relative tothe same variable x, such that RSn(x) c RSn'(x),their difference is the interpretation of the diffe-rence of their rank spaces: n - n' it is

called an abstraction (see Fig . 1) .

Figure 1. Equivalence (=k) links between QSkelements (k = 0, . . ., n + 1) ; elements are indexedby their ranks, i.e . denoted as qR(q) .

4 Operations on QS

4 .1 General Definitions and Properties

Def. 4.1.1: Internal operation .Let *k be a k-place operation on Z (k = 1, 2), ak-place operation on any QSn(x) is a function©k defined by:" ©1 : QS,(x) -~ QSn(x), b qj e Qsn(x),

©1 Ch. = In(* 1 &(gi))" ©2: [Qsn(x)] 2 -4QSn(x), b' gi, qj E [QSn(x)]2 ,

qj ©2 qj = InWgi) *2 R11(gj))

P.4.1.1: Closure .If Z is closed under *k, so is QSn(x) under ©k.

Def. 4.1.2: Generalizing operations on anyQS . Let any QSn(x), QSn'), QSn"(z), and anyqi E QSn(x), qj e QSn'(Y) ; external operationsare transformed into internal operations bychanging their operands into their equivalentsonto QSn.. (z) using a function c as defined byD.3 .3, that is (for ::= the rewriting sign) :" ©1 : QSr,(x) -~ QSn"(z), is changed into a func-tion QSn-(Z) -~ QSn°(z), by :

©1 gi : := ©1 c(gi)" ©2: QSn(x) x QSn'(y) --> QSn(z) is changedinto a function [QSn,.(z)]2 -4 QSn,(z) by:

(qi ©2 qj) : := (c(g) ©2 c(gj))with : c(%) = [qi],'%Qsn"(z)and

c(qj) =[qj]=n'/QSn"(z) .

zwW2WQ

On-1 q-n+1 q-n+2 - - q0 - - qn-2 qn-1

05% q-An+i - -I

qN

11111 INQSn+i q-n-tq-n q-n+1 _ _ _ q-1 q0 q1 _ - _ qn-1 qn qn+1

Oso

OS1

q0

q-~0 1

/11 N 0OS2 q-2

q_qO q

If. n < n" and n'< n" (transformations by c arerefinements), ©1 yields (n" - n + 1) solutionsand ©2 (n" - n + 1)(n" - n'+ 1) ; otherwise if:n" < n and n" < n' (transformations by c areabstractions) they yield only one solution.

P.4.1.2 : Morphism .An interpretation I is a homomorphism from (Z,

inv(gi 0 qi) = qi 0 qi, qi 0 qi = elm.

4 .5 Power and Root

D.4.5.1: An element power ([ .] .) another is theinterpretation of the product of both ranks :d qi , qj E QSn(x), [qi ]qi = I(Rgi)'R(qd)

P.4.5.1 : Properties of [.] ." The power is commutative and associative." 1(1) (element of rank 1) is its identity element :d qi, [qi ]I(1) = [I(1)]q` = qi

Since the existence of an inverse for all elementsis lacking (QS,(x), [.]-) are not groups but abe-lian monoids." Combination with inv :

[qi Iinv(gi) = inv([gi ]qi) = qm0 [qi ]qi" The norm is absorbent: `d qi, [qi ]qm = qm" Compatibility with®:

[qi ]qi = qi (9 . . .0 qi,

Rqj) rimes .

D.4.5.2: The root (j) of an element is the in-terpretation of the entire division of its rank bythe rank of the root :d qi E Qsn(x), d qi * qm,

qj

gk iff: R(qi) = R(gj) - R(qk)

qi - (gk,gt , ) ff:R(gi)=R(qj)-R(qk)+'L

= R(qj) - R(qk, ) + a

for ti, a E Z, i > 0, a < 0 and d(qk, qk.) = 1(closest neighbors of the exact result).

Table 1: Corresponding notions between

P.4.5.2 : Properties of j

" Identity element : b' qi,qi

" Self-inverse : d qi # qm,

qi =1(1)

qi" Absorption : b' qi # cj,,,

qm = qrn" Compatibility with the power.

qi

-over 0:

qi

" Distributivity : `d qi # qm,qj qj qj

-over ®:

qi ®qk = qi ®qk

qj qj qj

giogk = qi0 qk

4 .6 Operations Semantics

We can set some semantic equivalence between:(i) I and exponentiation, and (ii) the rankingfunction Rand logarithms (see Table 1). Withthis respect, a quality space has to some extendthe semantic value of a geometric progression,whereas the rank space stands for an arithmeticprogression (with ratio 1) . This is why we callmultiplication, division, power and root, opera-tions

on

QS

based

on

x , /, in Z,respectively .

the DuAl and exponentiation as a model.

alit Spaces Model Substitution Rulesvalues qi, qj, qk E QS,(x) Yi, Yj, Yk E 9t+Ranks zi =Nqi ) E RS,(x) c Z xi = 1ogayi E 9tInterpretation qi = Kzi )= I(R(gi )) Yi = axi

Specialelements

qminqm

01

0 :~ gmin1 qm

Addition qi ® gj = max(gi yi + y . = amax(xi, xi) +Subtraction qi 8gj = qi Yi - Y = ax,Mult . inverse inv(gi) = 1(-Nqi )) ( yi )-1 = axi inv( .)Multiplication qi ® qi = KR(g) + R(qj)) yi y . = axiax j = axi + xj x ::=Division qi 0 qj = KR(gi) - tqi)) -yi /y . = axi/axi = axi xj Q)

Power [qi lv= IMi)'N ')) (Yi )xl = (axi)xj = axi .xj [ .Root qj

qi

(yi )1/xj = axi /xi..= T

5 Two Examples of Application

5 .1 Analysis of a Non-Linear Model

We take from De Kleer and Brown (1984) theexample of the Cochin's law:

Q = CA

2P

(Eq. c 1)P

where Q is the flowrate through a valve, C thedischarge coefficient through the orifice of areaA, P the pressure across the valve and p themass density of the fluid. Since C, A, P and pare all positives, using the sign algebra-basedformalism yields [Q] = +. Instead of that, onewould like to assess the influence of the magni-tude of any of the variables C, A, p and P, onthe value of Q. That is, answering questionssuch as: "What if C is low, A small, P muchabove normal and p is more than very heavy ?" .Similarly : "keeping C, A, and P values constant,what would be the influence of p ?".Transforming the Cochin's expression into theDuAl formalism yields :

1(2)Q=C®A®(P®P)Op

since P ® P = P (idempotence), the expressionsimplifies to:

1(2)Q = C ®A®POp

(Eq. c2)Quality spaces we may choose (this is ourchoice) for the variables are:QS 1(A) = {small, regular, large)QS2(Q) = (very-small, small, OK, big,

very-big)QS2(C) = (very_low, low, normal, high,

very-high)QS2(P) = (m below, below, normal, above,

m_above)QS3(P) = [VVL, VL, L, M, H, VH, VVH}

Answering the first question is achieved aftersubstitution in Eq. c2 of known variables fortheir values :

1(2)

However to perform a consistent calculation,those values must be expressed in terms of theirequivalent in QS2(Q) using transformation func-tions (Def. 3.3) : c : q 1 i-4 c(q j ) = [qt]=2/QS2(Q).So, the initial values of known variables to beused in the calculus are given by:C = lowH c(low) = (small)A = smallH c(small) = (very-small, small }P = m-above i-4 c(m_above) = (verybig)p = VVHH c(VVH) = (very_big)

Since we have two values for A, we get twopossible solutions :(i) if A = very-small

1(2)Q = small ® very_ small ® very_ bigOvery_ big(ii) if A = small

1(2)Q = small ® small ®very_bigOvery_ bigHowever, since:(small ® very-small)

= (small (9 small)= very_small,

solutions (i) and (ii) sum up to only one solution:1(2)

Q = very_ small ®very_ bigovery_ big

1(2)=very_ small ®OK =very_small ®OK

that is : Q = verysmall, which answers theabove first question.To answer the second question, keeping theabove values for C, A, and P, let us calculate Qaccording to all the possible values of p bysolving the expression derived from instanciationof Eq. c2 by known values:

1(2)Q = low® small ® m_aboveOpusing equivalent values and simplifications asabove, it is rewritten as :

1(2)Q = very_ small ®very_bigOpSolutions are in table 2 . Although the initialmodel is non-linear, we get a linear response .

Q = low® small ®m_aboveOVVH

Table 2: Calculus of Q according to p (for C = low, A = small, P = m above) .

VVL VL L M H VH WHvery-small very-small small OK big v bi ve bi

Q OK OK OK,small

small small,ve small

very_small very-small

Table 3: Calculus of P from C = low, A = small, and values of p and Q in table 2 .

In order to check the stability of calculus withrespect to a transformation of the initial expres-sion, let us calculate P from the initial values ofC and A (i.e . C = low, A = small), and all theconfigurations of p and Q values expressed intable 2.The transformation is performed as follows :

1(2)Q=C®A®Pop

1(2)QO(C &A) = Pop

=* [QO(C ® A)]I(2) = Pop

=>P = p ®[QO(C (9 A)]I(2 ) .The calculus must lead to P = m-above . The re-sults are summarized in table 3, which showsthat stability is reasonably insured since all thesolutions cover the expected one.

5 .2 Relative Orders of Magnitude

The O(M) formalism provides a set of primitiverelations to compare quantities one to each other(Mavrovouniotis & Stephanopoulos, 1988) .The semantics of a relation is the comparison ofthe ratio of both quantities with respect to 1, sayx < y iff: x/y < 1 .We can compute this using the DuAl formalismby considering the ratio x/y as a single variabledepending on x and y values such that:x/y =xOy, and taking as a QS:

Q83(x/y ) = { },the primitive relations of O(M), ordered withrespect to strict equality == (x/y = q E QS3(x/y),means x q y) .For sake of clarity, let x and y quality spaces bethe same:QS3(x) = QS3(y) = {tiny, v_small, small,

normal, tall, v-tall, giant) .The respective values ofx/y are given in table 4.Applied to the counter current heat exchanger

example described in Mavrovouniotis andStephanopoulos (1988) we can derive conclu-sions consistent with O(M).The model is based on two equations :" Temperatures differences :DTH-DTI -DTC + DT2 = 0

(Eq. ol)" Energy balance :DTH.KH.FH = DTC.KC.FC

(Eq. o2)in which DTx are temperature differences (H andC stand for hot and cold flows, respectively),KH and KC are the molar-heats, FH and FC arethe molar flowrates .Given initial conditions:

DT2/DTl = (--),

let us derive : DT2/DTH, DTC/DTH, DTl/DTC,and FC/FH." DT2/DTH = DT2/DT1 . DTl/DTH

: := DT2/DTl ® DTl/DTH= (-

Table 4 : x/y = xOy where x/y has the O(M) primitive relations as codomain.

6 Conclusions and Perspectives

The DuAl formalism states that two elements ofsame rank are "equal", not because they are re-presented by the same label or their real magni-tude are equal, but because their distance withrespect to their own norm within some order areequal. This captures the intuitive notion ofassessment or preference scales . Let apple pie bean element of the quality space for cakes, androasted beef be an element for meat meals ;saying : "I like apple pie as much as roasted beefcomes from the fact they are both interpreted as"good meals" in my own scale of preference formeals, although they are different things.Pushing more, I may even say that an apple pieis greater than America's cup watched at the te-levision . This may seem unintuitive and so-mewhat scruffy, unless I add that if I had tochoose between both activities I would prefereating an apple pie than watching TV . Though,the idea is that apple pie, roasted beef, watchingthe America's cup, are objects ordered with res-pect to some common preference, utility, orassessment scale . This scale is the distance ofobjects to norms and all norms have the sameutility (zero) value of being a reference for somerelevant objects. Stating an equivalence betweentwo objects having the same interpretation in acommon quality space is thus the underlying ideaof our equivalence relation at level k.Therefore, reasoning about quantities, may alsolead to deal with ordinal instead of cardinal va-lues . That is why we resolved to choose the termquality space instead of quantity space to denotethe set of possible values for any variable .According to Forbus' definition (1984), a quan-tity space is a finite set of distinguishable valuesfor signs (taking values on (-1, 0, +1)), magni-tudes (taking values on 9t+v (0)) and finallynumbers, composed of both sign and magnitude(taking values on 9t) . In fact, quantity spaces arecollections of numerical values, eventually

interpreted as linguistic labels (e.g ., Davis,1987) . Therefore, operations of comparison andcombination on them are well defined, using theclassical algebraic operations or adaptation ofthem (as in the sign or interval algebras) . This isalso the case in Raiman's (1991) order of magni-tude reasoning (e.g . coarse values are sets ofnumbers) . If our QS values qc stand for somekind of magnitudes, they are not numbers sincethe sign component is lacking. More generally,Hayes' concept of quality spaces denotes the setof all possible values of a quantity, whatevertheir nature be . In those, as in our QS, may existnotions such as distance between values, tole-rance, non density, as well as functions to mapqualities with some linear order . DuAl is closerto this concept.As emphasized by Struss (1990), most of thework in qualitative reasoning has been directedtowards the qualitative interpretation ofquantita-tive equations (i.e . real-valued) . Our approach isa bit different . Starting from qualitative, local,poorly structured knowledge, our aim is to pro-vide a representation framework enabling one toderive unknown values from existing ones. TheDuAl approach came from difficulties encounte-red in modeling complex systems, such as eco-logical or biological, in which it is often useful tocompare and combine what is not comparable orcombinable . Human reasoning does so, butclassical models do not.In terms of applications, a preliminary approach,using mainly ad hoc combination tables andsome primitive operations, was proved useful toheuristic knowledge representation (Guerrin,1991) . In the current work, we put the emphasison operations. Although the task of representingknowledge in the form of equations is not easy,the use of operations with known properties gua-rantees sound calculations and provide verifica-tion possibilities, especially when many possiblevalues exist. However we think this will not re-frain heuristic representation if : (i) the semanticsof operations is clear, (ii) the modeler can check

z-->c(x)->

Y J. CW1

tiny>

tiny < < >> >> >> >>v small >> >> >>small >> >>normal =_ >tall >~ -

several combinations before choosing the ade-quate one, (iii) identification techniques are setup to help find the right model from partial adhoc tables . Verification of these aspects is aperspective . However, dealing with first prin-ciples equations, and comparison to other QRsystems, is also possible, as shown in section 5examples . Particularly, DuAl fits the general de-finition given by Williams (1991) of a qualitativealgebra, as an abstraction of a quantitative one(here integers arithmetic) . The DuAl approachwas also checked against Raiman's example(1991) of colliding masses, with consistent ana-lytical results except in one case . This short-coming comes from the lack of additive inverse(in the sense of our 9) and the brute force defi-nition of addition and subtraction, which is alimitation . In fact, our quality spaces, made ofpositive elements, are more likely to the Roughset (not closed under subtraction) of Raiman'sapproach, whereas the Small set (including zero)is lacking . A means to connect our Rough sets toSmall sets to improve calculation is anotherperspective .The refinement process for expanding a QS is,till now, based on refining its extremities. This isbecause greatest and least elements are very oftenconsidered as right and left open intervals res-pectively (e.g . any value above or below nume-rical thresholds, respectively) . We are currentlygeneralizing this notion of expanding a QS byrefining all the elements, be they extremities orintermediate elements . This may lead also to de-fine more gradual operations and certainly im-prove the modeling power of the formalism byimproving its flexibility .

Acknowledgements

Thanks to Waldir L. Roque (Federal Universityof Rio Grande do Sul, Brazil) and Paulo Salles(The University of Edinburgh, Scotland) fordiscussions and useful remarks about this work.Thanks also to referees ; their comments will behelpful to forthcoming improvements .

References

Davis, E. : Order of magnitude reasoning inqualitative differential equations, TR-312,Computer Science Dept., NY University,New York, 1987 .

De Kleer, J., Brown, J. : A qualitative physicsbased on confluences, Artificial Intelligence,1984, 24: 7-83 .

Forbus, K. : Qualitative Process Theory,Artificial Intelligence, 1984, 24: 85-168 .

Hayes, P. : The second naive physics manifesto,

in : J . Hobbs & R. Moore (Eds.), Formaltheories of the commonsense world, Ablex,Norwood, 1985, 71-90 .

Guerrin, F. : Qualitative reasoning about anecological process : interpretation in Hydro-ecology . Ecological Modelling, 1991, 59165-201 .

Kuipers, B. : Qualitative Simulation, ArtificialIntelligence, 1986, 29: 289-338 .

Mavrovouniotis, M.L., Stephanopoulos, G. :Formal order of magnitude reasoning inprocess engineering, Computers & ChemicalEngng, 1988, 12: 867-880.

Raiman, O. : Order of magnitude reasoning,Artificial Intelligence, 1991, 51 : 11-38 .

Struss, P. : Problems of interval-based qualitativereasoning, in : D Weld & de Kleer J.,Readings in Qualitative Reasoning aboutPhysical Systems, Morgan Kaufmann Publ.,San Mateo, CA, 1990, 288-305 .

Williams, B. : Qualitative analysis of MOScircuits, Artificial Intelligence, 1984, 24: 281-346.

Williams, B . : A theory of interactions : unifyingqualitative and quantitative algebraicreasoning, Artificial Intelligence, 1991, 51 :39-94 .